Class Boundaries Research Articles

MI-DABAN: A dual-attention-based adversarial network for motor imagery classification

The brain–computer interface (BCI) based on motor imagery electroencephalography (EEG) is widely used because of its convenience and safety. However, due to the distributional disparity between EEG signals, data from other subjects cannot be used directly to train a subject-specific classifier. For efficient use of the labeled data, domain transfer learning and adversarial learning are gradually applied to BCI classification tasks. While these methods improve classification performance, they only align globally and ignore task-specific class boundaries, which may lead to the blurring of features near the classification boundaries. Simultaneously, they employ fully shared generators to extract features, resulting in the loss of domain-specific information and the destruction of performance. To address these issues, we propose a novel dual-attention-based adversarial network for motor imagery classification (MI-DABAN). Our framework leverages multiple subjects’ knowledge to improve a single subject’s motor imagery classification performance by cleverly using a novel adversarial learning method and two unshared attention blocks. Specifically, without introducing additional domain discriminators, we iteratively maximize and minimize the output difference between the two classifiers to implement adversarial learning to ensure accurate domain alignment. Among them, maximization is used to identify easily confused samples near the decision boundary, and minimization is used to align the source and target domain distributions. Moreover, for the shallow features from source and target domains, we use two non-shared attention blocks to preserve domain-specific information, which can prevent the negative transfer of domain information and further improve the classification performance on test data. We conduct extensive experiments on two publicly available EEG datasets, namely BCI Competition IV Datasets 2a and 2b. The experiment results demonstrate our method’s effectiveness and superiority.

Component Identification and Prediction of Ancient Glass Products Based on Decision Tree Model and SVM Model

The study of the patterns of chemical composition of glass objects is an important research method of classifying ancient glass objects. In this paper , we firstly selected suitable indicators and sought classification boundary lines through a decision tree model to discover the classification pattern of ancient glass and verify its accuracy . Secondly, it was sub-classified and divided by means of hierarchical clustering and k-means mean clustering to realise the work of categorisation of known components . Finally an SVM model was built to obtain a confusion matrix map to achieve the classification prediction of unknown components.

A Study of Real-time Semantic Segmentation Performance Improvement in Unstructured Outdoor Environment

Semantic segmentation in autonomous driving for unstructured environments is challenging due to the presence of uneven terrains, unstructured class boundaries, irregular features and strong textures. Current off-road datasets exhibit difficulties like class imbalance and understanding of varying environmental topography. To overcome these issues, we propose a deep learning framework for semantic segmentation that involves a pooled class semantic segmentation with five classes. The evaluation of the framework is carried out on two off-road driving datasets, RUGD and TAS500. The results show that our proposed method achieves high accuracy and real-time performance.

Semi-supervised bidirectional alignment for Remote Sensing cross-domain scene classification

Remote sensing (RS) image scene classification has obtained increasing attention for its broad application prospects. Conventional fully-supervised approaches usually require a large amount of manually-labeled data. As more and more RS images becoming available, how to make full use of these unlabeled data is becoming an urgent topic. Semi-supervised learning, which uses a few labeled data to guide the self-training of numerous unlabeled data, is an intuitive strategy. However, it is hard to apply it to cross-dataset (i.e., cross-domain) scene classification due to the significant domain shift among different datasets. To this end, semi-supervised domain adaptation (SSDA), which can reduce the domain shift and further transfer knowledge from a fully-labeled RS scene dataset (source domain) to a limited-labeled RS scene dataset (target domain), would be a feasible solution. In this paper, we propose an SSDA method termed bidirectional sample-class alignment (BSCA) for RS cross-domain scene classification. BSCA consists of two alignment strategies, unsupervised alignment (UA) and supervised alignment (SA), both of which can contribute to decreasing domain shift. UA concentrates on reducing the distance of maximum mean discrepancy across domains, with no demand for class labels. In contrast, SA aims to achieve the distribution alignment both from source samples to the associate target class centers and from target samples to the associate source class centers, with awareness of their classes. To validate the effectiveness of the proposed method, extensive ablation, comparison, and visualization experiments are conducted on an RS-SSDA benchmark built upon four widely-used RS scene classification datasets. Experimental results indicate that in comparison with some state-of-the-art methods, our BSCA achieves the superior cross-domain classification performance with compact feature representation and low-entropy classification boundary. Our code will be available at https://github.com/hw2hwei/BSCA.

Continuous Indexing of Fibrosis (CIF): improving the assessment and classification of MPN patients

The grading of fibrosis in myeloproliferative neoplasms (MPN) is an important component of disease classification, prognostication and monitoring. However, current fibrosis grading systems are only semi-quantitative and fail to fully capture sample heterogeneity. To improve the quantitation of reticulin fibrosis, we developed a machine learning approach using bone marrow trephine (BMT) samples (n = 107) from patients diagnosed with MPN or a reactive marrow. The resulting Continuous Indexing of Fibrosis (CIF) enhances the detection and monitoring of fibrosis within BMTs, and aids MPN subtyping. When combined with megakaryocyte feature analysis, CIF discriminates between the frequently challenging differential diagnosis of essential thrombocythemia (ET) and pre-fibrotic myelofibrosis with high predictive accuracy [area under the curve = 0.94]. CIF also shows promise in the identification of MPN patients at risk of disease progression; analysis of samples from 35 patients diagnosed with ET and enrolled in the Primary Thrombocythemia-1 trial identified features predictive of post-ET myelofibrosis (area under the curve = 0.77). In addition to these clinical applications, automated analysis of fibrosis has clear potential to further refine disease classification boundaries and inform future studies of the micro-environmental factors driving disease initiation and progression in MPN and other stem cell disorders.

Reducing Data Complexity Using Autoencoders With Class-Informed Loss Functions.

Available data in machine learning applications is becoming increasingly complex, due to higher dimensionality and difficult classes. There exists a wide variety of approaches to measuring complexity of labeled data, according to class overlap, separability or boundary shapes, as well as group morphology. Many techniques can transform the data in order to find better features, but few focus on specifically reducing data complexity. Most data transformation methods mainly treat the dimensionality aspect, leaving aside the available information within class labels which can be useful when classes are somehow complex. This paper proposes an autoencoder-based approach to complexity reduction, using class labels in order to inform the loss function about the adequacy of the generated variables. This leads to three different new feature learners, Scorer, Skaler and Slicer. They are based on Fisher's discriminant ratio, the Kullback-Leibler divergence and least-squares support vector machines, respectively. They can be applied as a preprocessing stage for a binary classification problem. A thorough experimentation across a collection of 27 datasets and a range of complexity and classification metrics shows that class-informed autoencoders perform better than 4 other popular unsupervised feature extraction techniques, especially when the final objective is using the data for a classification task.

Supervised Contrastive Learning-Based Domain Adaptation Network for Intelligent Unsupervised Fault Diagnosis of Rolling Bearing

Fault diagnosis of rolling bearing is essential to guarantee production efficiency and avoid catastrophic accidents. Domain adaptation is emerging as a critical technology for the intelligent fault diagnosis of rolling bearing. Most existing solutions learn domain-invariant features by statistical moment matching, adversarial training, or fusing two algorithms. However, these domain adaptation methodologies overemphasized learning domain-invariant features and ignored the generalization of classification performance on the target domain, which leads to inevitable misclassification. To address this issue, we propose a supervised contrastive learning-based domain adaptation network (SCLDAN) for cross-domain fault diagnosis of the rolling bearing in this paper. The SCLDAN develops a 1-D convolutional residual network to learn the raw signal features and employs the maximum mean discrepancy loss to achieve global domain alignment. In addition, a novel supervised contrastive learning approach is proposed, where a supervised contrastive loss and a mutual information loss are established to learn the class-specific information and improve the reliability of target prediction labels. Thus, the ambiguous data samples residing near the class boundaries of the target domain can be accurately identified, and the diagnosis accuracy is significantly improved. Extensive experiments on two experimental scenarios demonstrate the effectiveness of the proposed method.

Resonant transport of fusion alpha particles in quasisymmetric stellarators

In modern, highly optimized stellarator configurations where prompt fusion alpha particle losses from the plasma core are absent, alpha particles can still be lost due to stochastic orbits which result in delayed losses. One mechanism leading to stochastic orbits are changes in the particle trapping class during drift motion along the contours of the parallel adiabatic invariant J ∥ leading to jumps in J ∥ when crossing class boundaries. Another mechanism, which is of main interest here, is the resonance between particle drift and bounce motion (drift-orbit resonance). The first mechanism affects mainly trapped particles near the trapped-passing boundary in the phase space of quasi-symmetric and quasi-isodynamic devices, and can be minimized there by aligning local magnetic field maxima on a given flux surface. The second mechanism may affect a broader range in the trapped particle domain where contours of J ∥ still remain closed. Drift-orbit resonances modify the topology of orbits leading to island-like structures on Poincaré plots where these islands may overlap thus leading to the stochastic transport. In this report, we study this stochastization mechanism in quasi-symmetric stellarator configurations with help of the Hamiltonian drift-kinetic code NEO-RT as well as orbit classification and direct computation of fusion alpha losses within the symplectic orbit following code SIMPLE. The width and overlap of resonances in phase-space is studied using Hamiltonian perturbation theory. Based on optimized reactor configurations we assess if this approach can be used as a fast metric for fusion alpha losses in stellarator optimization.

Generalizing Correspondence Analysis for Applications in Machine Learning.

Correspondence analysis (CA) is a multivariate statistical tool used to visualize and interpret data dependencies by finding maximally correlated embeddings of pairs of random variables. CA has found applications in fields ranging from epidemiology to social sciences. However, current methods for CA do not scale to large, high-dimensional datasets. In this paper, we provide a novel interpretation of CA in terms of an information-theoretic quantity called the principal inertia components. We show that estimating the principal inertia components, which consists in solving a functional optimization problem over the space of finite variance functions of two random variable, is equivalent to performing CA. We then leverage this insight to design algorithms to perform CA at scale. Specifically, we demonstrate how the principal inertia components can be reliably approximated from data using deep neural networks. Finally, we show how the maximally correlated embeddings of pairs of random variables in CA further play a central role in several learning problems including multi-view and multi-modal learning methods and visualization of classification boundaries.

Subspace-based minority oversampling for imbalance classification

In pattern classification, the class imbalance problem always occurs when the number of observations in some classes is significantly different from that of other categories, which leads to the learning bias in the classifiers. One possible solution to this problem is to re-balance the training set by over-sampling the minority class. However, over-samplings always push the classification boundaries to the majority part, thus the recall increases while the precision decreases. To avoid this situation and better handle the class imbalance problem, this paper proposes a new over-sampling method, namely Subspace-based Minority Over-Sampling (abbr. SMO). This approach considers that each category of samples is formed by common and unique characteristics, and such characteristics can be extracted by subspace. To obtain the balanced data, the common part is over-sampled for more accurately depicting the minority, and the unique part can be expanded by some generative methods. The balanced data are obtained by restoring the generated products of the subspace to the original space. The experimental results demonstrate that the SMO has the ability to model complex data distributions and outperforms both classical and newly designed over-sampling algorithms. Also, SMO can be used to generate simple images, and the generation results of MNIST can be clearly identified by both human vision and machine vision.

Dynamic identification of key nodes in active distribution network for operation optimisation requirements

AbstractAccurate and dynamic identification of key nodes that can efficiently improve the safety and economy of active distribution network (ADN) is of great significance for optimising the operation level of ADN with time‐varying characteristics. In this study, a dynamic identification method for the key nodes of ADN operation optimisation requirements is proposed. Firstly, considering the influence of nodes on ADN operation performance, the importance of nodes in the network structure, and the controllable potential of nodes, a multi‐angle key node identification index system is constructed. Secondly, considering the influence of the ADN operation state on node importance, the classification boundary of ADN operation state was set to determine the operation state of the ADN at different times. Thirdly, the order relation analysis method and variation coefficient method were used to comprehensively evaluate the key nodes of the ADN under different operating states. Finally, the technique for order preference by similarity to an ideal solution is used to sort the comprehensive evaluation results of key nodes to achieve the effective identification of key nodes. The effectiveness and feasibility of the proposed method are verified on 33‐bus and 118‐bus test systems considering the time sequence variation in distributed generation and load.

Vibration Image Representations for Fault Diagnosis of Rotating Machines: A Review

Rotating machine vibration signals typically represent a large collection of responses from various sources in a machine, along with some background noise. This makes it challenging to precisely utilise the collected vibration signals for machine fault diagnosis. Much of the research in this area has focused on computing certain features of the original vibration signal in the time domain, frequency domain, and time–frequency domain, which can sufficiently describe the signal in essence. Yet, computing useful features from noisy fault signals, including measurement errors, needs expert prior knowledge and human labour. The past two decades have seen rapid developments in the application of feature-learning or representation-learning techniques that can automatically learn representations of time series vibration datasets to address this problem. These include supervised learning techniques with known data classes and unsupervised learning or clustering techniques with data classes or class boundaries that are not obtainable. More recent developments in the field of computer vision have led to a renewed interest in transforming the 1D time series vibration signal into a 2D image, which can often offer discriminative descriptions of vibration signals. Several forms of features can be learned from the vibration images, including shape, colour, texture, pixel intensity, etc. Given its high performance in fault diagnosis, the image representation of vibration signals is receiving growing attention from researchers. In this paper, we review the works associated with vibration image representation-based fault detection and diagnosis for rotating machines in order to chart the progress in this field. We present the first comprehensive survey of this topic by summarising and categorising existing vibration image representation techniques based on their characteristics and the processing domain of the vibration signal. In addition, we also analyse the application of these techniques in rotating machine fault detection and classification. Finally, we briefly outline future research directions based on the reviewed works.

Fast prototype selection algorithm based on adjacent neighbourhood and boundary approximation

The unceasing increase of data quantity severely limits the wide application of mature classification algorithms due to the unacceptable execution time and the insufficient memory. How to fast incrementally obtain high decision reference set and adapt to incremental data environment is urgently needed in incremental environments, large dataset, etc. This paper proposes a novel prototype selection algorithm by integrating the strategies between condensing method and editing method. To an unlearned pattern, this algorithm extends the references scope from its single nearest neighbour to its k nearest neighbourhood that can expand the judgment information to obtain its detailed neighbour relationship. Then a pattern was determined whether it is a prototype using its neighbour relationship and classification boundary asymptotically strategy. To maintain the higher reference set, this algorithm periodically updates those prototypes that locates in the non-boundary zone or is long-time unlearned. The empirical study shows that this algorithm obtains the smaller and higher boundary prototypes without decreasing classification accuracy and reduction rate than the compared algorithms.

A unified deep semi-supervised graph learning scheme based on nodes re-weighting and manifold regularization

In recent years, semi-supervised learning on graphs has gained importance in many fields and applications. The goal is to use both partially labeled data (labeled examples) and a large amount of unlabeled data to build more effective predictive models. Deep Graph Neural Networks (GNNs) are very useful in both unsupervised and semi-supervised learning problems. As a special class of GNNs, Graph Convolutional Networks (GCNs) aim to obtain data representation through graph-based node smoothing and layer-wise neural network transformations. However, GCNs have some weaknesses when applied to semi-supervised graph learning: (1) it ignores the manifold structure implicitly encoded by the graph; (2) it uses a fixed neighborhood graph and focuses only on the convolution of a graph, but pays little attention to graph construction; (3) it rarely considers the problem of topological imbalance.To overcome the above shortcomings, in this paper, we propose a novel semi-supervised learning method called Re-weight Nodes and Graph Learning Convolutional Network with Manifold Regularization (ReNode-GLCNMR). Our proposed method simultaneously integrates graph learning and graph convolution into a unified network architecture, which also enforces label smoothing through an unsupervised loss term. At the same time, it addresses the problem of imbalance in graph topology by adaptively reweighting the influence of labeled nodes based on their distances to the class boundaries. Experiments on 8 benchmark datasets show that ReNode-GLCNMR significantly outperforms the state-of-the-art semi-supervised GNN methods.11The code is available at https://github.com/BiJingjun/ReNode-GLCNMR.

Improving the repeatability of deep learning models with Monte Carlo dropout

The integration of artificial intelligence into clinical workflows requires reliable and robust models. Repeatability is a key attribute of model robustness. Ideal repeatable models output predictions without variation during independent tests carried out under similar conditions. However, slight variations, though not ideal, may be unavoidable and acceptable in practice. During model development and evaluation, much attention is given to classification performance while model repeatability is rarely assessed, leading to the development of models that are unusable in clinical practice. In this work, we evaluate the repeatability of four model types (binary classification, multi-class classification, ordinal classification, and regression) on images that were acquired from the same patient during the same visit. We study the each model’s performance on four medical image classification tasks from public and private datasets: knee osteoarthritis, cervical cancer screening, breast density estimation, and retinopathy of prematurity. Repeatability is measured and compared on ResNet and DenseNet architectures. Moreover, we assess the impact of sampling Monte Carlo dropout predictions at test time on classification performance and repeatability. Leveraging Monte Carlo predictions significantly increases repeatability, in particular at the class boundaries, for all tasks on the binary, multi-class, and ordinal models leading to an average reduction of the 95% limits of agreement by 16% points and of the class disagreement rate by 7% points. The classification accuracy improves in most settings along with the repeatability. Our results suggest that beyond about 20 Monte Carlo iterations, there is no further gain in repeatability. In addition to the higher test-retest agreement, Monte Carlo predictions are better calibrated which leads to output probabilities reflecting more accurately the true likelihood of being correctly classified.

A three-way classification with fuzzy decision trees

This study is concerned with the design of a three-way classification mechanism realized through combing fuzzy decision trees and expressing uncertainty associated with classification results. The fuzzy decision tree used in this study is constructed through the generalization of the commonly used decision trees. The notion of three-way decision model first proposed for the interpretation of rules generated in rough set approximation has been widely used in many fields. This study proposes an efficient way to flag data with high level of uncertainty with the classification realized by fuzzy decision trees, which is the capability that commonly used fuzzy decision trees do not have. The data identified in this way are left to users’ judgments or more advanced classification techniques. The developed mechanism is formed as a two-stage construct where a fuzzy decision tree is built by generalizing the Boolean classification boundaries of a pre-constructed decision tree using fuzzy sets and then determining the level of uncertainty to identify instances to be rejected due to a lack of sufficient confidence in their belongingness. The rejected instances that are difficult to process are classified as non-commitment cases and left to some further analyses. The rejection quality of the developed three-way classifier is quantified in terms of the classification accuracy and rejection coefficient. We also elaborate on striking a sound tradeoff between these two performance indicators. Experimental studies demonstrate that the developed mechanism could effectively improve the classification accuracy at the cost of a small proportion of the rejected instances and achieve better performance in comparison with other three-way decision models when generating a three-way decision output.

Unbinding Gentility: Women Making Music in the Nineteenth-Century South by Candace Bailey

Reviewed by: Unbinding Gentility: Women Making Music in the Nineteenth-Century South by Candace Bailey Julia Nitz Unbinding Gentility: Women Making Music in the Nineteenth-Century South. By Candace Bailey. Music in American Life. (Urbana, Chicago, and Springfield: University of Illinois Press, 2021. Pp. xx, 292. Paper, $30.00, ISBN 978-0-252-08574-1; cloth, $125.00, ISBN 978-0-252-04375-8.) This volume is admirable in its ambition to capture the role southern amateur women played in American musical history, or, as the author puts it, to tell the "history of women in the nineteenth-century U.S. South . . . through the medium of music" (p. 4). Candace Bailey critiques past studies that considered women's musical artifacts and practices "ephemera," a misplaced notion shared by musicologists and cultural historians alike, and she strives to overcome traditional, male-centric narratives "that prejudice composer and title over circulation and performance" (p. 2). Bailey's book "encompasses music's transmission, education, circulation, and repertory in order to understand music's meaning in women's culture of the South" (p. 3). Thus, she recharts southern musical traditions, putting special emphasis on the role of "scientific music" in women's cultural engagement across class and racial boundaries (p. 15). Key to Bailey's study is the location of "women's musical practices in the performance of gentility" (p. 5). She convincingly illustrates how musical practices during this era were intricately tied to social hierarchies and distinctions, showing how different strata of society used music as cultural capital. Tracing these developments from the antebellum period to Reconstruction, she explores the effect of the dissolution of the South's social hierarchy on music making. [End Page 761] The material consulted in this study is extensive, with a focus on music read from notation in over fifteen hundred binder's volumes. Bailey also examines "numerous nineteenth-century magazines and journals, newspapers, letters and diaries, official records, and many similar documents" (p. 2). She uses these materials to assemble ethnographic readings and microhistories that collectively reveal the kaleidoscopic, multifaceted southern music-scape of the mid-nineteenth century. Bailey proceeds chronologically in five sections: Parts 1, 2, and 3 cover the antebellum period, Part 4 explores the Civil War, and Part 5 turns to Reconstruction. Each part is divided into several chapters, and each chapter offers a distinct study. Parts 1, 2, and 3 provide an in-depth exploration of the plethora of antebellum musical performers. It illustrates the ubiquity of music in society and highlights how economic growth in the 1840s and 1850s opened new avenues to gentility for the middle and poorer classes. This broadening of the sphere of gentility is evidenced by a more widespread "literate music practice" than hitherto assumed (p. 15). Part 2 focuses on prominent genres of the period and the accompanying repertory used to signify gentility, including a chapter on operatic tunes. While Parts 1 and 2 mostly provide nuance to existing studies of antebellum women's music, Part 3 reframes repertory history. It deals with the role of a genre known as scientific music in women's musical education and practice. Bailey argues that its pervasiveness was an important marker of social status. Part 4, dealing with the Civil War years, focuses on social changes and how they triggered both retroactive and progressive musical practices. In these chapters especially, Bailey shows the importance of adding music to a reading of the parlor as a space for the construction of gender, class, and ethnicity. Part 5, on women musicians in the Reconstruction era, identifies the rise of women as professional performers and the changes in repertory toward classical music. Bailey convincingly illustrates how women musicians "redefined women's sphere and influence after the war by linking domestic competence and control with a public persona" (pp. 171–72). Bailey reconfigures traditional assumptions about the musical contributions of southern women by exploring the interrelation between music making and gentility in the U.S. South. Her approach is more inclusive than seen in previous studies, charting the musical practices of white women and women of color of all social classes. The representation of Black women in her study remains sketchy by necessity. As...

Intelligent Bearing Fault Diagnosis Based on Open Set Convolutional Neural Network

Traditional data-driven intelligent fault diagnosis methods have been successfully developed under the closed set assumption (CSA). CSA-based fault diagnosis assumes that the fault types in the test set are consistent with that in the training set, which can achieve high accuracy, but this is generally not valid in real-world industrial applications where the collection of data in industrial applications is often limited. As it is unrealistic to assume that the training set will cover all fault types, the application of the fault classifier may fail when the test set contains unknown fault types because the probability of input samples belonging to unknown types cannot be obtained. To solve the problem of how unknown fault types may be accurately identified, this paper further studies the open set assumption (OSA) fault diagnosis. We propose an open set convolutional neural network (OS-CNN) method and apply our OS-CNN model to an improved OpenMax method as a deep network to accurately detect unknown fault types. The overall performance was significantly improved as our OS-CNN model was able to effectively tighten the boundary of known classes and limit the open-space risk for the OpenMax method based on distance modeling. The overall effectiveness of the proposed method was verified by experimental studies based on four different bearing datasets. Compared with state-of-the-art OSA fault diagnosis method, our method cannot only realize the correct classification of the known fault classes, but it can also accurately detect the unknown fault classes.

Model and Training Method of the Resilient Image Classifier Considering Faults, Concept Drift, and Adversarial Attacks

Modern trainable image recognition models are vulnerable to different types of perturbations; hence, the development of resilient intelligent algorithms for safety-critical applications remains a relevant concern to reduce the impact of perturbation on model performance. This paper proposes a model and training method for a resilient image classifier capable of efficiently functioning despite various faults, adversarial attacks, and concept drifts. The proposed model has a multi-section structure with a hierarchy of optimized class prototypes and hyperspherical class boundaries, which provides adaptive computation, perturbation absorption, and graceful degradation. The proposed training method entails the application of a complex loss function assembled from its constituent parts in a particular way depending on the result of perturbation detection and the presence of new labeled and unlabeled data. The training method implements principles of self-knowledge distillation, the compactness maximization of class distribution and the interclass gap, the compression of feature representations, and consistency regularization. Consistency regularization makes it possible to utilize both labeled and unlabeled data to obtain a robust model and implement continuous adaptation. Experiments are performed on the publicly available CIFAR-10 and CIFAR-100 datasets using model backbones based on modules ResBlocks from the ResNet50 architecture and Swin transformer blocks. It is experimentally proven that the proposed prototype-based classifier head is characterized by a higher level of robustness and adaptability in comparison with the dense layer-based classifier head. It is also shown that multi-section structure and self-knowledge distillation feature conserve resources when processing simple samples under normal conditions and increase computational costs to improve the reliability of decisions when exposed to perturbations.

Source-free unsupervised domain adaptation with maintaining model balance and diversity

Source-free unsupervised domain adaptation (SFUDA) uses the knowledge learned from the source domain pre-trained model to perform the task on the target domain without directly accessing the source domain. However, the source domain pre-trained model is biased towards the source domain but deviates from the target domain, so its pseudo-label on the target domain tends to have a large noise. Thus, we should avoid using noisy pseudo-labels to train the model. In this article, we propose a novel kind of source-free unsupervised domain adaptation with maintaining model balance and diversity (SFMBD), which designs a target domain-specific classifier whose classification boundary is far from the high-density area of the target domain feature distribution. In addition, we keep the model balance and promote the model diversity while maintaining its ability to discriminate the target domain. Experimental evaluation of multiple benchmark datasets illustrates the effectiveness of our proposed approach on SFUDA.